[rwth-security] MINJA sneak attack poisons AI models for other chatbot users